Chaos Under Control

Happy Monday!

Last week, Stripe unveiled a new AI foundation model for payments, revealing how the payment giant is applying artificial intelligence to detect fraud and optimize payment systems. But beneath the headlines lies a fascinating paradox about how businesses are integrating AI: the most successful enterprise applications of generative AI aren't actually letting AI be fully generative.

The Meta Trend: Making the Unpredictable Predictable

For all its promise, generative AI has a fundamental challenge for businesses: it's probabilistic by nature. Unlike traditional software that delivers the same result every time given the same inputs, large language models introduce randomness and variation. This non-deterministic behavior, while the source of AI's creative power, poses significant problems for enterprises that need consistent, reliable results.

Before we jump in, I want to clarify what we mean by "probabilistic“ and “deterministic” in the first place. LLM’s are essentially language prediction machines; context goes in, all of the words and paragraphs get converted into numbers, and then some vector math happens to determine the most likely output for the given input.

This is great for creative writing or brainstorming, but what happens when LLM’s “hallucinate” in their responses? Hallucinations are essentially when low probability responses appear in the final output. If you say “the sky is” and hit enter, the majority of the time the LLM will respond with “blue.” There are smaller percentages that it will respond with other matching phrases like “cloudy” or “rainy” or “stormy,” and the context of the entire question usually helps shift the final response in the right direction. But there is also a non-zero chance that the LLM responds with “cloudy with a chance of meatballs” because the phrase appeared one time in a children’s book.

As someone intelligent once told me, hallucinations are funny when you’re asking for a cupcake recipe. They aren’t funny when you’re flying an airplane. For enterprise companies incorporating AI, they can’t even take a small chance that the LLM response will fall into the low percentage of “unlikely” answers. They need to create guardrails to ensure that answers are predictable without feeling robotic.

The emerging solution isn't abandoning AI's probabilistic nature but rather building deterministic frameworks around it. Companies like Stripe, Goldman Sachs, and JPMorgan aren't just deploying raw LLMs. They're actually creating hybrid systems that pair AI's generative capabilities with structural guardrails that ensure predictable business outcomes.

Pattern Recognition: The Deterministic Revolution

Four patterns highlight how enterprise companies are building deterministic guardrails around AI applications:

1. Stripe's AI-Enhanced, Rules-Based Approach: Stripe hasn't abandoned its deterministic payment processing backbone while embracing AI. Instead, they've integrated AI into specific nodes of existing workflows where probability is acceptable or where multiple safety checks can validate outputs.

   As one Stripe executive acknowledged in a panel discussion, "generative AI gives you new capabilities and also new challenges. So, the inherent uncertainty is there but we need to figure out how to de-risk it." The solution involves short feedback loops and validation checkpoints to ensure AI-enhanced systems deliver consistently.

2. Goldman Sachs' Platform with Built-in Guardrails: Goldman Sachs has developed a comprehensive approach to AI deployment through its GS AI Platform, which offers multiple state-of-the-art AI models, but with Goldman's guardrails firmly in place.

   The bank's platform isn't just allowing access to powerful models, it offers controlled deployment with security and governance as foundational guardrails. This allows Goldman to maintain deterministic business processes while leveraging AI's probabilistic strengths in a protected environment.

3. Microsoft's Agent Flows for Deterministic Workflow Automation: Microsoft recently announced "agent flows" in Copilot Studio, a system explicitly designed to address the deterministic/probabilistic divide. The company describes agent flows as combining "outcome-driven autonomous agents with deterministic agent flows."

   These flows "automate deterministic workflows" while being "enhanced with powerful AI actions." The key insight is that unlike "purely dynamic autonomous orchestration," deterministic workflows "leverage predefined action sequences" to ensure tasks "always run consistently and rapidly."

4. UPS's Human-AI Approval Workflows: UPS has implemented a generative AI system called MeRA for customer service, but with a crucial twist: it automates responses to customer emails while giving human agents final approval. This human-in-the-loop approach reduced agent handle time by 50% during testing while ensuring every response meets company standards.

   This hybrid workflow represents a crucial pattern; rather than replacing deterministic processes with probabilistic AI, companies are finding ways to leverage AI while maintaining ultimate control over outcomes.

The Contrarian Take: The Real AI Revolution Is in the Guardrails

The conventional narrative suggests that AI's value comes from unleashing its creative, probabilistic nature. But the real story emerging from enterprise AI deployments tells a different tale: the true revolution isn't in the models themselves, but in the guardrails, approval flows, and validation systems companies are building around them.

Businesses aren't actually embracing the uncertainty of traditional LLM’s. The most successful enterprise AI implementations don't simply deploy foundation models; they create deterministic frameworks that harness AI's probabilistic power for specific purposes while ensuring consistent business outcomes.

This distinction matters because it suggests that competitive advantage in AI won't come merely from access to powerful models (which are increasingly commoditized), but from the systems companies build to make these models reliably productive within their specific business contexts. The wrapper will matter just as much as what’s inside powering these systems.

Practical Implications

For Enterprises:

• Focus on building deterministic frameworks around probabilistic AI models, not just deploying the models themselves

• Identify specific workflow nodes where AI can operate with controlled uncertainty

• Implement validation and approval systems to ensure consistency in business-critical AI applications

• Consider human-in-the-loop approaches that combine AI efficiency with human judgment

For AI Teams:

• Develop expertise in both AI capabilities and process engineering to create effective guardrails

• Build monitoring systems that detect when AI outputs drift beyond acceptable parameters

• Design workflows where deterministic processes handle critical path operations while AI enhances capabilities

• Create feedback mechanisms that continually improve AI outputs while maintaining consistency

For Business Leaders:

• Recognize that the competitive advantage lies not in the AI models themselves but in how they're integrated into business processes

• Invest in the "connective tissue" between AI capabilities and deterministic business requirements

• Focus on end-to-end outcomes rather than AI capabilities in isolation

• Build cross-functional teams that combine AI expertise with deep business process knowledge

The future of enterprise AI likely isn't fully autonomous systems making independent decisions. Rather, it's sophisticated hybrid architectures where probabilistic AI enhances fundamentally deterministic business processes.

Companies that master this balancing act of leveraging AI's creative potential while ensuring predictable outcomes will gain significant advantages over those that either avoid AI entirely, or deploy it without proper guardrails.

The question is how businesses will build deterministic frameworks to make the power that LLM’s provide safe, reliable, and consistently valuable.

In motion,
Justin Wright

1. Superintelligent AI Agents for Scientific Discovery (FutureHouse)

2. UiPath Agentic Automation (UiPath)

3. Nvidia Teams With Peachtree Corners Smart City to Expand Capabilities (AI Business)

4. Ottawa Hospital uses AI ambient voice capture to reduce physician burnout (VentureBeat)

5. Singapore’s Vision for AI Safety Bridges the US-China Divide (Wired)